CHAPTER 3. Single-view Geometry. 1. Consequences of Projection

Transcription

1 CHAPTER 3 Single-view Geometry When we open an eye or take a photograph, we see only a flattened, two-dimensional projection of the physical underlying scene. The consequences are numerous and startling. Size relationships are distorted, right angles are suddenly wrong, and parallel lines now intersect. Objects that were once apart are now overlapping in the image, so that some parts are not visible. At first glance, it may seem that the imaging process has corrupted a perfectly reasonable, well-ordered scene into a chaotic jumble of colors and textures. However, with careful scene representation that accounts for perspective projection, we can tease out the structure of the physical world. In this chapter, we summarize the consequences of the imaging process and how to mathematically model perspective projection. We will also show how to recover an estimate of the ground plane, which allows us to recover some of the relations among objects in the scene. 1. Consequences of Projection The photograph is a projection of the 3D scene onto a 2D image plane. Most cameras use a perspective projection. Instead of recording the 3D position of objects (X,, Z), we observe their projection onto the 2D image plane at (u,v). In this projection, a set of parallel lines in 3D will intersect at a single point, called the vanishing point. A set of parallel planes in 3D intersect in a single line, called the vanishing line. The vanishing line of the ground plane, called the horizon, is of particular interest because many objects are oriented according to the ground and gravity. See Figure 3.1 for an illustration of vanishing points and lines. Another major consequence of projection is occlusion. Because light does not pass through most objects, only the nearest object along a ray is visible. Take a glance at the world around you. Every object is occluding something, and most are partially occluded themselves. In some ways,

2 CHAPTER 3. SINGLE-VIEW GEOMETR Vanishing line Vertical vanishing point (at infinity) Vanishing point Vanishing point FIGURE 3.1. Illustration of vanishing points and lines. Lines that are parallel in 3D intersect at a point in the image, called the vanishing point. Each set of parallel lines that is parallel to a plane will intersect on the vanishing line of that plane. The vanishing line of the ground plane is called the horizon. If a set of parallel lines is also parallel to the image plane, it will intersect at infinity. Three finite orthogonal vanishing points can be used to estimate the intrinsic camera parameters. Figure from [37].. Optical Center (u 0, v 0 ) f Z. u v u p = v. Camera Center (t x, t y, t z ). X P = Z FIGURE 3.2. Illustration of pinhole camera model. The camera, with focal length f and optical center (u 0,v 0 ), projects from a 3D point in the world to a 2D point on the image plane. occlusion simplifies the scene interpretation. Imagine if you could see every object for miles in a single projection! In other ways, occlusion makes scene interpretation more difficult by hiding parts and disturbing silhouettes. 2. Perspective Projection with Pinhole Camera: 3D to 2D We often assume a pinhole camera to model projection from 3D to 2D, as shown in Figure 3.2. Under this model, light passes from the object through a small pinhole onto a sensor. The pinhole model ignores lens distortion and other non-linear effects, modeling the camera in terms of its intrinsic parameters (focal length f, optical center (u 0,v 0 ), pixel aspect ratio α, and skew s) and extrinsic parameters (3D rotation R and translation t). Although the image plane is physically behind the pinhole, as shown in the figure, it is sometimes convenient to pretend that the image 16

3 3 3D MEASUREMENT FROM A 2D IMAGE plane is in front, with focal length f, so that the image is not inverted. If we assume no rotation (R is identity), and camera translation of (0,0,0), with no skew and unit aspect ratio, we can use the properties of similar triangles to show that u u 0 = f X Z and v v 0 = f Z. We can more easily write the projection as a system of linear equations in homogeneous coordinates. These homogeneous coordinates add a scale coordinate to the Cartesian coordinates, making them convenient for representing rays (as in projection) and direction to infinitely distant points (e.g., where 3D parallel lines intersect). To convert from Cartesian to homogeneous coordinates, append a value of 1 (e.g., (u,v) (u,v,1) and (X,,Z) (X,,Z,1)). To convert from homogeneous to Cartesian coordinates, divide by the last coordinate (e.g., (u, v, w) (u/w, v/w)). Under homogeneous coordinates, our model of pinhole projection from 3D world coordinates (X,,Z) to image coordinates (u,v) is written as follows: (3.1) w u X w v = K[R t] Z or w u w v = f s u 0 0 α f v 0 w w r 11 r 12 r 13 t x r 21 r 22 r 23 t y r 31 r 32 r 33 t z Note that the intrinsic parameter matrix K has only five parameters and that the rotation and translation each have three parameters, so that there are 11 parameters in total. Even though the rotation matrix has nine elements, each element can be computed from the three angles of rotation. We commonly assume unit aspect ratio and zero skew as a good approximation for most modern cameras. Sometimes, it is also convenient to define world coordinates according to the camera position and orientation, yielding the simplified (3.2) w u w v = f 0 u 0 0 f v 0 X. w Z From this equation, we get the same result as we found using similar triangles in Figure 3.2. In Cartesian coordinates, u = f X Z + u 0 and v = f Z + v D Measurement from a 2D Image As illustrated in Figure 3.3, an infinite number of 3D geometrical configurations could produce the same 2D image, if only photographed from the correct perspective. Mathematically, therefore, we have no way to recover 3D points or measurements from a single image. Fortunately, because our world is so structured, we can often make good estimates. For example, if we know how the camera is rotated with respect to the ground plane and vertical direction, we can recover the relative 3D heights of objects from their 2D positions and heights. X Z 1. 17

4 CHAPTER 3. SINGLE-VIEW GEOMETR Original Image Interpretation 1: Painting on Ground Interpretation 2: Floating Objects Interpretation 3: Man in Field FIGURE 3.3. Original image and novel views under three different 3D interpretations. Each of these projects back into the input image from the original viewpoint. Although any image could be produced by any of an infinite number of 3D geometrical configurations, very few are plausible. Our main challenge is to learn the structure of the world to determine the most likely solution. Depending on the application, the world coordinates can be encoded using either three orthogonal vanishing points or the projection matrix, leading to different approaches to 3D measurement. When performing estimation from a single image, it is helpful to think in terms of the horizon line (the vanishing line of the ground plane) and the vertical vanishing point. For example, look at the 3D scene and corresponding image in Figure 3.4. Suppose the camera is level with the ground, so that the vertical vanishing point is at infinity and the horizon is a line through row v h. Then, if a grounded object is as tall as the camera is high, the top of the object will be projected to v h as well. We can use this and basic trigonometry to show that o c = v t v b v h v b, where o is the 3D height at the top of the object (defining the ground at = 0), c is the camera height, and v t and v b are the top and bottom positions of the object in the image. This relationship is explored in more detail in Hoiem et al. [104] and used to aid in object recognition, a topic we will discuss in a later chapter. 18

5 3 3D MEASUREMENT FROM A 2D IMAGE v h v t v b c =0 o o c v v t h v v b b FIGURE 3.4. The measure of man. Even from one image, we can measure the object s height, relative to the camera height, if we can see where it contacts the ground. FIGURE 3.5. This figure by Steve Seitz illustrates the use of the cross-ratio to make world measurements from an image. As shown by Criminisi et al. [38], the cross ratio invariant can be used to relate the 3D heights of objects under more general conditions. The cross ratio of 4 collinear points (e.g., p 3 p 1 p 4 p 2 p 3 p 2 p 4 p 1, for any collinear points p 1..4 ) does not change under projective transformations. If one of those points is a vanishing point, then its 3D position is at infinity, simplifying the ratio. Using the vertical vanishing point, we can recover the relative heights of objects in the scene (see Figure 3.5 for an illustration). 19

6 CHAPTER 3. SINGLE-VIEW GEOMETR We have a special case when the camera is upright and not slanted so that the vertical vanishing point is at infinity in the image plane. In this case, v Z =, so we have t b c b = v t v b v h v b, where t is the height of the top of the object, b is the height at the bottom, c is the height of the camera, and v t, v b, and v h, are the image row coordinates of the object and horizon. 4. Automatic Estimation of Vanishing Points Recall that all lines with the same 3D orientation will converge to a single point in an image, called a vanishing point. Likewise, sets of planes converge to a vanishing line in the image plane. If a set of parallel lines is also parallel to a particular plane, then their vanishing point will lie on the vanishing line of the plane. If we can recover vanishing points, we can use them to make a good guess at the 3D orientations of lines in the scene. As we will see, if we can recover a triplet of vanishing points that correspond to orthogonal sets of lines, we can solve for the focal length and optical center of the camera. Typically, two of the orthogonal vanishing points will be on the horizon and the third will be in the vertical direction. If desired, we can solve for the rotation matrix that sets the orthogonal directions as the X,, and Z directions. Most vanishing point detection algorithms work by detecting lines and clustering them into groups that are nearly intersecting. If we suspect that the three dominant vanishing directions will be mutually orthogonal, e.g., as in architectural scenes, we can constrain the triplet of vanishing points to provide reasonable camera parameters. However, there are a few challenges that make the problem of recovering vanishing points robustly much more tricky than it sounds. In the scene, many lines, such as those on a tree s branches, point in unusual 3D orientations. Nearly all pairs of lines will intersect in the image plane, even if they are not parallel in 3D, leading to many spurious candidates for vanishing points. Further, inevitable small errors in line localization cause sets of lines that really are parallel in 3D not to intersect at a single point. Finally, an image may contain many lines, and it is difficult to quickly find the most promising set of vanishing points. Basic algorithm to estimate three mutually orthogonal vanishing points: 1. Detect straight line segments in an image, represented by center (u i,v i ) and angle θ i in radians. Ignore lines with fewer than L pixels (e.g., L=20), because these are less likely to be correct detections or to have large angular error. Kosecka and Zhang [123] describe an effective line detection algorithm (see Derek Hoiem s software page for an implementation, currently at 20

7 5 SUMMAR OF KE CONCEPTS 2. Find intersection points of all pairs of lines to create a set of vanishing point candidates. It s easiest to represent points in homogeneous coordinates, p i = [w i u i w i v i w i ] T, and lines in the form l i = [a i b i c i ] T using the line equation ax + by + c = 0. The intersection of lines l i and l j is given by their cross product: p i j = l i l j. If the lines are parallel, then w i j = 0. Likewise, the line formed by two points p i and p j is given by their cross product: l i j = p i p j. 3. Compute a score s j for each vanishing point candidate (u j,v j ): s j = i l i exp ( α i θ i 2σ 2 ) where l i is the length and θ i is the orientation (in radians) of line segment i, α i is the angle from the center of the line segment (u i,v i ) to (u j,v j ) in radians, and σ is a scale parameters (e.g., σ = 0.1). In the distance between angles, note that 2π = 0 = 2π. 4. Choose the triplet with the highest total score that also leads to reasonable camera parameters. For a given triplet, p i, p j, p k, the total score is s i jk = s i +s j +s k. Orthogonality constraints specify: ( K 1 p i ) T ( K 1 p j ) = 0; ( K 1 p i ) T ( K 1 p k ) = 0; and ( K 1 p j ) T ( K 1 p k ) = 0, where K is the 3x3 matrix in Eq. 3.2 and the points are in homogeneous coordinates. With three equations and three unknowns, we can solve for u 0, v 0, and f. We consider (u 0, v 0 ) within the image bounds and f > 0 to be reasonable parameters. The algorithm described above is a simplified version of the methods in Rother [189] and Hedau et al. [94]. Many other algorithms are possible, such as [123, 226, 11], that improve efficiency, handle vanishing points that are not mutually orthogonal, incorporate new priors, or have different ways of scoring the candidates. Implementations from Tardif [226] and Barinova et al. [11] are available online. Sometimes, for complicated or non-architectural scenes, example-based methods [239, 104] may provide better estimates of perspective than vanishing point based methods. 5. Summary of Key Concepts The camera projects 3D points onto a 2D imaging plane. In consequence, parallel 3D lines converge to a vanishing point on the image plane, size relationships and angles between lines are distorted, and nearer objects occlude further objects along a ray. The true sizes, positions, and orientations of objects and surfaces can often be recovered by modeling the camera, typically with a pinhole camera model. The reader should understand and memorize the pinhole projection equation (Eq. 3.1) and the notation of the intrinsic and extrinsic camera matrices. To recover 3D size and position in an image, we need to know the distance and orientation between the camera and the supporting 3D plane, typically through vanishing point estimation or detection of known objects., 21